What is Amazon?

In Why It's Hard To Escape Amazon's Long Reach, Paris Martineau and Louise Matsakis have compiled an amazingly long list of businesses that exist inside Amazon's big tent. After it went up, they had to keep updating it as people pointed out businesses they'd missed. In most of those businesses, Amazon's competitors are at a huge disadvantage:

While its retail business is the most visible to consumers, the cloud computing arm, Amazon Web Services, is the cash cow. AWS has significantly higher profit margins than other parts of the company. In the third quarter, Amazon generated $3.7 billion in operating income (before taxes). More than half of the total, $2.1 billon, came from AWS, on just 12 percent of Amazon’s total revenue. Amazon can use its cloud cash to subsidize the goods it ships to customers, helping to undercut retail competitors who don’t have similar adjunct revenue streams.

In the mid-50s my father wrote a textbook, Organisation of retail distribution, with a second edition in the mid-60s. He would have been fascinated by Amazon. I've written about Amazon from many different viewpoints, including storage as a service, and anti-trust, so I'm fascinated with Amazon, too. Now, when you put recent posts by two different writers together, an extraordinarily interesting picture emerges, not just of Amazon but of the risks inherent to the "friction-free" nature of the Web:

• Zack Kanter's What is Amazon? is easily the most insightful thing I've ever read about Amazon. It starts by examining how Walmart's "slow AI" transformed retail, continues by describing how Amazon transformed Walmart's "slow AI" into one better suited to the Internet, and ends up with a discussion of how Amazon's "slow AI" seems recently to have made a fundamental mistake.
• Izabella Kaminska's series Amazon (sub)Prime? and Amazon (sub)Prime - Part II provides the deep dive to go with Kanter's big picture, looking in detail into one of the many symptoms of the "slow AI's" apparent mistake.

Below the fold, a long meditation on these posts.

So how come Zack Kanter knows so much about Amazon?

I have sold to and bought from Amazon in about as many ways as one person can; I built an auto parts brand that sold thousands of SKUs [stock keeping units] to Amazon as a vendor (both stocking and drop ship) and as a marketplace seller (both "seller-fulfilled" and Fulfillment By Amazon, or FBA), before selling the company to a private equity fund in 2018. And I am now the founder and CEO of a startup called Stedi (a modern EDI platform, if you're familiar with EDI) that runs on Amazon Web Services; we automate transactions like purchase orders and invoices between brands and retailers.

You really need to take the time right now to go read Kanter's "short book" then come back here at this link. Trust me, it will be time well spent. But, if you can't be bothered, here is enough of a summary so you can understand the rest.

Kanter starts by describing the goals of Walmart's "slow AI":

1. "a wide assortment of good quality merchandise"
2. offered "at the lowest possible prices"
3. backed by "guaranteed satisfaction" and "friendly, knowledgeable service"
4. available during "convenient hours" with "free parking" and "a pleasant shopping experience"
5. all within the largest, most convenient possible store size and location permitted by local economics

As the "slow AI" was seeking these goals it was constrained by the limited shelf space in each store. Optimizing the profit to be made from the limited shelf space meant that Walmart's mechandisers had to find the best products, negotiate the lowest prices, ensure that the vendor could keep up with Walmart's demands, and so on. Walmart built a huge organization that was extraordinarily good at doing this. But then:

Jeff Bezos had a big realization in 1994: the world of retail had, up until then, been a world where the most important thing was optimizing limited shelf space in service of satisfying the customer - but that world was about to change drastically. The advent of the internet - of online shopping - meant that an online retailer had infinite shelf space.

The goals of Amazon's "slow AI" were simpler:

1. a vast selection
2. delivered fast
3. at the lowest possible prices
4. backed by guaranteed satisfaction

With infinite shelf space and search functionality:

the new formula became simpler: the more SKUs it added, the more items would be discovered by customers; the more items that customers discovered, the more items they would buy. In this world of infinite shelf space, it wasn't the quality of the selection that mattered - it was pure quantity. And with this insight, Amazon did not need to be nearly as good - let alone better - than Walmart at Walmart's masterful game of vendor and SKU selection. Amazon just needed to be faster at aggregating SKUs - and therefore faster at onboarding vendors.

The juggernaut started rolling:

Amazon systematically removed friction from the seller onboarding workflow, doing seemingly small things like eliminating the UPC code requirement that would serve as a barrier for newer, less established sellers. All of these small changes started to add up, and Amazon became the fastest way for a company to start selling online. Customers began to associate Amazon with selection, and Amazon became the de facto storefront for the fledgling world of online commerce. ... Amazon's SKU aggregation juggernaut was running an unbound search for customer value nationwide, while Walmart's army of finely-tuned retailer gatekeepers was still running a bounded search in local geographies.

Kanter continues with a fascinating description of how Amazon, faced with a series of constraints on its growth, each time solved them with a platform that others, outside the company, could use. The obvious example is AWS, but there are many.

Thanks to these platforms, the juggernaut kept running and Amazon became this Internet-scale collection of SKUs. Kanter writes:

you have to understand that things get really weird when you run an unbounded search at internet-scale. When you remove "normal" constraints imposed by the physical world, the scale can get so massive that all of the normal approaches start to break down.

Amazon's version of this was the sheer size and diversity of its collection of SKUs:

Amazon would never be able to effectively curate such a sprawling array of product categories. It isn't particularly good at merchandising to start with, and, even if it were, it could never build a large enough army of merchandisers to curate such a massive selection. Instead, Amazon relies on a ranking algorithm that heavily weights product reviews and sales velocity. The more reviews a product has and the more units it sells, the higher it climbs in rankings. Of course, this creates a positive feedback loop: the more a product is exposed to customers, the more it sells; the more it sells, the more reviews it gets, and the higher it climbs in rankings, starting the loop all over again. (Yes, this is a gross oversimplification of Amazon's extraordinarily complex ranking algorithm)

This creates a big problem for Amazon's customers, who want the latest and greatest products, and for its sellers, who want to develop and sell exciting new items. Failure to satisfy these demands would put Amazon's ecommerce dominance at risk.

Amazon answered this problem in typical fashion: with a platform. Amazon Advertising allowed sellers to feature 'Sponsored Products' - paid ads that appear at the top of search results. Sponsored Products solved three problems at once: new product discovery for the customers, new product introductions for the sellers, and, as an added bonus, pure gross margin revenue for Amazon - to the tune of $8 billion annually.

In other words, since the curation problem was too big for Amazon to solve, they made a for-pay curation platform to solve it. Introducing "Sponsored Products" in this way is very similar to what Google did to curate search results:

Google allows companies to bid on search terms, and displays paid content at the top of its search results in the same blue font used for unpaid content. (For example, a candy maker might bid on the term "Christmas candy" so that its ads pop up when someone searches for those words.) Google identifies ads in its search results with an icon below the link.

When this topic came up on Dave Farber's IP list, my friend Chuck McManis, who once ran a Google competitor, weighed in with a typically informative response about the result:

On the search page, Google's bread and butter so to speak, for a 'highly contested' search (that is what search engine marketeers call a search query that can generate lucrative ad clicks) such as 'best credit card' or 'lowest home mortgage', there are many web browser window configurations that show few, if any organic search engine results at all!

As I wrote at the time:

In other words, for searches that are profitable, Google has moved all the results it thinks are relevant off the first page and replaced them with results that people have paid to put there. Which is pretty much the definition of "evil" in the famous "don't be evil" slogan notoriously dropped in 2015. I'm pretty sure that no-one at executive level in Google thought that building a paid-search engine was a good idea, but the internal logic of the "slow AI" they built forced them into doing just that.

Naturally, "Sponsored Products" caused the same thing to happen to Amazon:

The problem with Sponsored Products is that sponsored listings are not actually good for customers - they are good for sellers; more specifically, they are good for sellers who are good at advertising, and bad for everyone else. Paid digital advertising is a very specific skill set; the odds that the brand with the best product also happens to employ the best digital marketing staff or agency is extraordinarily low. Further, the ability to buy the top slot in search results favors products with the highest gross margin - hence the highest bidder - not the products that would best satisfy customers.

The issue is compounded by the fact that the average customer is unable to tell the difference between an "organic" search result and a sponsored product. The top four results in an Amazon search are now occupied by sponsored listings, which means that the average Amazon customer is disproportionately likely to be purchasing a sponsored product. And since the sponsored listings favor high-margin products pushed by savvy digital marketers, it is highly unlikely that Amazon's customer is buying the optimal product that the market could provide.

To be sure, very poor products get rated poorly and are weeded out quickly, but, by and large, sponsored listings drag the average quality of products sold closer to mediocrity, and further from greatness. That's bad.

What happens when you order one of these mediocre, high-margin products? Amazon ships it to you in two days, or maybe even next day, or in some cases the same day! How is this even possible?

This the question behind Izabella Kaminska's deep dive. Some of the products with rapid delivery are Amazon's own. Preventing Amazon both owning the platform and selling through it is a key aspect of Senator Elizabeth Warren's anti-trust policy; it isn't a good idea for a company to compete with its own customers. But that's a topic for a different post. The other products with rapid delivery come from vendors using "Fulfilled By Amazon" (FBA); they are stocked in and shipped from Amazon's warehouses. There are two ways to identify products in this system. The first is the FNSKU:

Unless you make your money from selling stuff on Amazon, chances are you won't have heard of an FNSKU. The acronym stands for Fulfilment Network Stock Keeping Unit and represents a location identifier for products sitting in Amazon warehouses. This, to all intents and purposes, equates to an Amazon barcode.

The other is the manufacturer's barcode, the one you'd see in a brick-and-mortar store. Amazon's systems discourage vendors from using the FNSKU:

Not using an FNSKU is appealing for sellers. It means products sourced from manufacturers do not have to be relabelled, ensuring they can be sent into Amazon's network directly, saving time and money. Sellers who have chosen to be fulfilled by Amazon otherwise add an additional logistical layer into their operations if they have to relabel the goods independently.

Using manufacture bar codes also means products are more likely to qualify for Amazon Prime classification, pushing them higher up the search rankings.

Amazon's explanation for why it prefers manufacturer's barcodes is:

If multiple sellers have inventory with the same manufacturer barcode, Amazon may fulfil orders using products with that barcode when those products are closest to the customer. This happens regardless of which seller actually receives a customer's order. We use this process to facilitate faster delivery.

In other words, products using an FNSKU will ship from the warehouse where the vendor stocked them, whereas products using a manufacturer's barcode will ship from a warehouse where a vendor selling the same product has stocked them. In theory this "commingling" is good for the vendors, getting them faster delivery at lower inventory cost, and good for the customers, who get faster delivery and lower prices.

At Internet scale there's no way Amazon can verify that products carrying the same manufacturer's barcode are actually the same product. So, inevitably, in some cases they aren't. Kaminska reports:

For the whole thing to work seamlessly, the underlying inventory across the entire system must be genuinely equivalent, and bear all the same properties. For that to be case, somebody has to be willing to police the quality of goods entering the system.

Sellers we spoke to said Amazon seems strangely reluctant to step up on this front, preferring to trust that what's on the label is what's inside - possibly because of the costs involved.

Some sellers believe the lackadaisical approach has now exposed the entire network to contamination by inferior or fake goods: an unscrupulous vendor can pass off a copycat good as the genuine article by applying a manufacturer's barcode, which is easy to do.

This in turn has created a quality lottery for customers who purchase from commingled inventory (often without realising it).

It gets worse:

because Amazon is still obliged to return unsold inventory to suppliers on request, sellers say this creates an incentive for opportunists to send in fake, or low quality, goods into commingled inventory just to receive higher quality goods in return. What proportion of quality goods they receive back depends on how contaminated the particular product pool is, but for many the arbitrage opportunity is worth a punt.

A further arbitrage relates to products bearing manufacturer warranties. As a rule, warranties are either dated from the date of manufacture or the date of sale. In the event of the former, commingled pools can comprise of a huge range of warranty durations, from entirely expired to brand-new. Since customers often don't check warranties until it's too late, these discrepancies often go unnoticed.

Just as Facebook claims it can police its network, Amazon claims it can handle the problem:

Amazon says it has the means to track the provenance of disputed goods even if sourced from commingled stock. It adds that in such cases it fronts the refund to the customer directly before attempting to recoup costs from the bad actor responsible. If that fails, it pursues further action against the bad actor - whether that's blocking the account, litigation or law enforcement.

But sellers insist that in many instances the company's response time is far too slow when it comes to disabling offending accounts or protecting their IP. It's also highly discretionary, with further legal action only being taken when it is expedient for Amazon to do so.

The vendors are much smaller than Amazon, and dependent upon Amazon's sales channel. So it is easy, if short-sighted, for Amazon to insist that they must deal with the problem themselves:

One US-based seller of security products told us:

Amazon is demanding sellers who suffer from this problem that the sellers - who have no way of actually rectifying the issue unless they de-intermingle [de-commingle] and start issuing their own SKUs which is against Amazon's policy - offer a plan of action to explain how they are going to fix all the customer complaints they are getting about the products Amazon is sending,

He added that if they don't offer a plan, Amazon simply kicks them off the platform anyway, as the presence of bad ratings is bad for everyone. The recent Amazon vendor purge is testament to that (more on that in our follow-up post).

Another seller said that the process of identifying the bad seller who supplied his customer is not only long-winded but typically involves placing a "test buy" themselves, putting the onus on them to prove the problem exists, which may or may not be replicable.

After this examination of the role of fake products in the Fulfilled By Amazon channel, in Part II, Kaminska discusses their role in Amazon Marketplace:

Fakes can also sneak in through the company's Marketplace platform. On Marketplace, Amazon only acts as a commission-charging matching agent, with sellers fulfilling their own orders and using their own warehouses, meaning there's little to no quality control or supervision on their part.

As usual, Amazon's way of dealing with this problem was to create a platform, the review and rating system, by which outsiders could solve it for them:

But reviews on the site cannot always be trusted, thus Amazon's largely immutable review and ratings system hasn't entirely eliminated the information asymmetry that creates the famous Akerlof-ian lemon problem — in which good products are underpriced to compensate for the risk of customers unwittingly buying lemons.

Customers have learnt to offset the risk of delays and returns by demanding significant discounts on products purchased via third parties. That discount, alongside the cost and uncertainty of running arduous returns policies, has transferred a huge amount of risk to the remaining honest sellers on Marketplace.

It may not be just FBA and Marketplace vendors that are at risk from fake products:

To what degree Amazon's own products (ie those it sources from wholesale suppliers directly) are vulnerable to being mixed in with third-party products, however, is unclear. All we know from the spokesman is that Amazon has been commingling its own products since 2013 in the UK. Whether or not there has been an increase in the number of subquality goods being sold under Amazon's name can only be determined by a holistic review of comments for Amazon's own service which, as far as we know, is impossible for a user to do.

Last month Amazon abruptly purged many long-standing vendors of Amazon's own products. Ostensibly, this was to push them into the Marketplace program, which generates better margins for Amazon. But:

Other evidence, however, suggests the vendor purge was as much to do with rooting out bad actors as it was about managing costs. For example, a week or so after the vendor purge was announced, DigiDay UK was among those reporting that Amazon had decided to walk back the decision to terminate purchase orders on the proviso vendors signed up to Amazon's Brand Registry enrolment system. This is a service that allows brand owners to protect their IP by ensuring those who sell their products on Amazon must have permission to do so. Except, suspensions didn't just impact brand owners or licensees.
As DigiDay put it:

The emphasis on Brand Registry suggests that Amazon’s Vendor Central purge was a move to eliminate vendors that not only aren’t profitable for Amazon to manage, but in some cases are also low quality, selling counterfeit goods or branded products without authorisation.

This, however, is understandably problematic for legitimate resellers of branded products.

Kaminska concludes:

it's fair to assume that the more good actors opt out of commingling the more they increase costs for Amazon and leave the online store itself on the hook for the subquality products entering its system from the remaining bad actors. ... the more trustworthy a retailer's supplier network is, the leaner and more efficient it can operate and, ultimately, the more cost competitive it can be.

What Amazon investors need to establish is to what degree Amazon's model of instant and affordable fulfilment under the Prime umbrella has been indirectly dependent on drawing on similar efficiencies without the same regard for quality control. And, in that context, to what degree its competitive pricing (and consequent ability to outcompete conventional retail) is more the product of allowing any old seller with any old ware to compete alongside those with superior standards — generating the lemon discount phenomenon — than a core efficiency in its model.

Because if the answer is a lot, the more we see Amazon taking measures to nip the counterfeiting problem in the bud, the more likely the retailer is to bifurcate into a closed online retail system (similar to conventional retail with trusted supply chains, white labelled goods and other customer-facing ops) and an open-ended eBay-equivalent (albeit without an auction mechanism to regulate prices).

Just as "Sponsored Products" give the advantage to sellers who exploit the buyer, commingling gives the advantage to unscrupulous sellers who exploit both more scrupulous sellers and the buyers, who get fake products or expired warranties.

The fundamental problem is one that affects all Internet scale platforms. It is the same problem that means Facebook can't prevent the bad guys using their platform to spread hate and interfere in elections, can't prevent third-parties leaking their user's data, and won't remove content that obviously violates their stated policies. It is the same problem that means the bad guys can censor YouTube with bogus DMCA takedowns, spread anti-vaxxer propaganda, and target children with disturbing fake Peppa Pig videos. It is the same problem that means the Google Play store contains a whole lot of malware, including from governments. It is the same problem that means scholarly communication is collapsing under attack from predatory publishers.

The problem is that Internet scale means "things get really weird" and human intervention to de-weird them is ineffective. Why is human de-weirding ineffective? For three related reasons:

• Even though the FAANGs are huge companies, there are a lot more bad guys out there trying to subvert their platforms than the companies have employees. Just as an example, as of last December there were 35,587 full-time Facebook employees. But last week, Researchers unearth 74 Facebook cybercrime groups with 385,000 members. Ten bad guys per employee in just this one example, which is not unique:
  

  Friday’s report comes a year after journalist Brian Krebs reported Facebook had deleted almost 120 groups with more than 300,000 members total after Krebs provided documentation they were flagrantly promoting a host of illicit activities on the social media network’s platform.

• The financial and other incentives for the bad guys are large and immediate. In the short term, and even as we see with Facebook the medium term, the financial penalty for not de-weirding their platform is insignificant.
• Because the short-term penalties for tolerating weirdness are insignificant, and the cost of even marginally effective efforts to de-weird the platform large - they require hiring large numbers of humans - the platforms only mount token de-weirding efforts. The problem being fixed isn't the weirdness, it is the publicity about the lack of de-weirding efforts. Note that Facebook depends upon outsiders like Krebs and Talos to detect bad guys, despite the searches involved being trivial:
  

  Indeed, less than two minutes of searching on Facebook turned up groups that appeared to offer the same services. One group called Carding Secured offered an array of services related to stolen payment-card data.

In each instance of the fundamental problem, the result is negative externalities. Fossil fuel companies' massive profits are based on avoiding the costs carbon emissions impose on current, and even more so future, society. The Sackler's massive profits are based on avoiding the costs opioid addiction imposes on current, and even more so future, society. Similarly, the platforms' massive profits are based on avoiding the costs that are imposed on current, and even more so future, society by their inability to cope with the way things get really weird when you run an unbounded search at internet-scale. Costs such as science's loss of credibility, a generation of traumatized, video-addicted children, and the collapse of democracy.